A0442
Title: Toward a unified ranking method for variable importance in regression models
Authors: Tsung-Chi Cheng - National Chengchi University (Taiwan) [presenting]
Abstract: Under various application analyses, evaluating the relative importance of individual explanatory variables in the regression model holds substantial practical significance and meaningfulness. Among many measurements of relative importance proposed in the literature, no single method can be universally regarded as the best. The dominance analysis (DA) framework assesses each variable's contribution to model fit across all subset models in a comprehensive and hierarchical manner. However, DA does not always yield a definitive ranking of variable importance. By applying the Condorcet winner criterion to the results based on conditional dominance within the DA framework, simulation studies for both linear and logistic regression models demonstrate an effective approach to ranking variable importance. This enhancement offers valuable improvements in the practical application of relative importance measurement.
